The same is true with observations of black ravens. If we see a black raven in a random sample of ravens, it is informative. It is just one data point, but it is part of a sample that can answer our questions. But the same black raven tells us nothing about our two raven questions if it is encountered in a sample of black things; there is no way to use such a sample to answer either question. The role of procedures is fundamental; an observation is only evidence if it is embedded in the right kind of procedure. I think this is a very general fact about evidence and confirmation; Hempel was wrong to think that generalizations are always confirmed by observations of their instances. There is only confirmation (or support) if the underlying procedure was of the right kind. (Interestingly, this does not apply in the case of deductive relationships. A black raven refutes the hypothesis that no ravens are black, regardless of the procedure behind the observation. But deduction, as always, is special.)
That concludes my sketch of a solution to the ravens problem. It is a more elaborate version of the idea, discussed in chapter 3, that "order of observation" is important (Horwich 198z). But what is important is not order, but procedures.
I turn now to the grue problem (section 3.4). This is harder, because I believe that "the grue problem" actually combines several different problems together (including the very difficult problem of simplicity). But I will present part of an answer.
Let us continue thinking about inferences made from samples, using statistical methods. These methods can be very powerful, but they can only be used when some assumptions hold about the testing situation. One situation where these methods cannot be used, in their simple forms, is when the act of observing or collecting the data changes the particular objects being observed, in a way that is relevant to the question being asked. In some cases we might overcome the problem by taking into account the effects of our data collection and compensating for this fact. But special measures of some kind will be needed.
Now consider Goodman and his emeralds. Again, the philosophical literature has chosen a bad example here, but suppose we are making inferences about all emeralds by observing a random sample. This method would encounter a problem if the act of collecting or observing individual emeralds changed their color. In such a case, a simple extrapolation from the color of the sample to the color of the unobserved emeralds would be un reliable. This problem is obvious. But there is a less obvious connection between this case and the grue problem.
First, I should remind you that a grue object is not one that changes color at some special date. I am not trying to solve the grue problem by objecting to emeralds changing their color, or anything like that. A grue object is one that was first observed before zoi o and is green, or was not observed before zoio and is blue. With that point clear, think about a sample of grue emeralds that we might have collected.
To keep things simple, suppose that all our previously observed emeralds are in the sample. So we have a big pile of emeralds, all of which are grue. The act of putting them in the sample did not physically change them, but something related is going on. If those particular emeralds had not been observed before zoio, they would not have been grue. After all, those emeralds are green, and anything green that was never observed before zoio does not count as grue. So the grueness of an object depends, in an odd conceptual way, on whether or not the object has been observed before a certain date. Putting it loosely, the emeralds in the sample were affected, with respect to their grueness, by the fact that they have been observed before now. But that means we cannot extrapolate grueness from the sampled emeralds to the unsampled ones. We cannot do the extrapolation because the observation process has interfered-in an odd way-with the characteristics of the objects in the sample. This problem does not appear if we want to extrapolate greenness from a sample of emeralds; it only appears if we want to extrapolate grueness.
The grue problem (or this aspect of the grue problem) is a strange philosophical version of a familiar problem in statistical methodology in science. It is akin to what would be called a confounding variable problem. In a way, Goodman's term "grue" turns observation (or sampling) itself into a confounding variable. Frank Jackson (1975) proposed a solution to Goodman's problem of roughly this kind, but without tying the solution to statistical methods or the idea of a confounding variable. I follow up this idea in more detail in Godfrey-Smith (forthcoming). The idea I want to emphasize here is, again, the importance of focusing on procedures in thinking about evidence.
Further Reading
On Bayesianism, I find Colin Howson and Peter Urbach's Scientific Reasoning: The Bayesian Approach (1993) very helpful, though the book does not seem to be popular within the Bayesian camp. John Earman, Bayes or Bust? (1992.), is for the technically minded. Skyrms, Choice and Chance (zooo), is a classic introduction to probability and induction. Michael Resnik's Choices (1987) is a particularly helpful introduction to decision theory, subjective probability, and Dutch books.
Farman (1992., chap. 7) and Kircher (1993, chap. 7) discuss and defend eliminative inference. (Neither of these is easy reading.) For one of the elaborate attempts to make sense of the scientific preference for simple theories, see Forster and Sober 1994.
The view that I call "procedural naturalism" amalgamates ideas from various sources. Reichenbach's main discussion is in Experience and Prediction (193 8) and is presented more accessibly in The Rise of Scientific Philosophy (1951). Alvin Goldman's two big books Epistemology and Cognition (1986) and Knowledge in a Social World (1999) give a general treatment of epistemological questions emphasizing the reliability of methods, rules, and procedures. The technically minded might be interested in a recent area of work that could be seen as contributing to a procedural naturalist approach to explanatory inference. This is work on inference about causal structure in networks of interacting factors (Pearl zooo; Spirtes, Glymour, and Sheines 1993).
15.1 A Muddy Paste?
We now reach the end of the tour. The tour has covered nearly a century of argument about science, and it has visited some fairly extreme climates and landscapes along the way. I will finish the book by trying to tie together some of the various threads, hints, insights, and pieces of the puzzle that have emerged in the preceding chapters. In particular, I will connect three ideas: empiricism, naturalism, and scientific realism. These three "isms" have each been explored and, in some form, defended. The harder question is whether they can be combined into a package that makes sense as a whole. We can't just declare that we are all "empiricist naturalist realists" or "naturalistic realist empiricists" and consider the job done.
When I wrote the proposal for this book, publishers sent it out for comments. One anonymous reviewer reacted against the idea that at the end we would have a happy three-way marriage of empiricism, naturalism, and scientific realism. The reviewer saw these as three ideas that could each be defended fairly well individually but which do not go well together. There are conflicts between them, or at least between some of the pairs. To make a good case for scientific realism, for example, requires being opposed to some central ideas in the empiricist tradition. So the reviewer predicted that when the last chapter of the book tried to put the three ideas together, the result would be a "muddy paste."
This is a good image. We start with three sharp and distinctive colors, three different big ideas about science, but when we try to put them together, we get a mess. Or so the reviewer predicted. Despite this vivid warning, I will indeed try to put the three together in this chapter. Readers can decide for themselves whether the result is mud.
15.2 The Apparent Tensions
Empiricism traditionally holds that our source of knowledge about the world is experience. Naturalism holds that we can only hope to resolve philosophical problems (including epistemological problems) by approaching them within a scientific picture of ourselves and our place in the universe. Scientific realism holds that science can reasonably aim to describe the real structure of the world, including its unobservable structur
e. So why can't we believe all three of these at once? Where is the problem?
Much of the problem comes from the side of empiricism. I have several times summarized empiricism as the view that our only source of knowledge is experience. But this is, of course, a vague and indefinite idea, more a starting point than a philosophical position. When people have tried to fill out this idea, the result has often been a view with troublesome consequences.
Traditional empiricism was often beguiled by a picture of the mind as shut in behind a "veil of ideas," or sensations. If all we have access to is our sensory experience, what chance do we have of forming justified beliefs about what lies beyond the veil? The most extreme forms of empiricism have denied that it even makes sense to talk or think about what might lie beyond experience. And even from a less extreme empiricist point of view, it can be hard to see how experience itself could support a hypothesis about structures lying behind experience. Hence the temptation to see science as concerned only with patterns in experience itself, or patterns in the observable domain.
In recent years the tension between scientific realism and empiricism has often been debated under the heading "the underdetermination of theory by evidence." Empiricists argue that there will always be a range of alternative theories compatible with all our evidence. So we can never have good empirical grounds for choosing one of these theories over others and regarding it as representing how the world really is. If we have no empirical grounds for such a choice, then we have no grounds at all.
So much for empiricism and scientific realism. The other possible tensions are not so bad, but they are still worth discussing. First, the relation between empiricism and naturalism is not always harmonious, because empiricist philosophies have often had a foundationalist structure. For many empiricists, given that we only have direct access to our ideas and experiences, we must begin from that starting point when developing a philosophical theory of knowledge. But according to naturalism, the idea of "starting within the circle of our ideas and then working our way out" is a bad mistake.
Some have thought there is a different kind of tension between naturalism and empiricism. Sociologists of science (and others in neighboring fields) have held that the empiricist tradition in philosophy of science has been shown to be a collection of myths. If we look at how science actually works, we do not find experience acting as a neutral "arbiter" of theoretical disputes, in the way imagined by empiricism. Arguments about the theoryladenness of observation (section 10.3) are often used to make this point. I tried to defuse most of these arguments, but they do continue to be influential.
The last possibility is tension between naturalism and scientific realism. Here we find less of a problem. Indeed, it is hard to be a naturalistic philosopher without taking science seriously as a description of the world; that suggests that naturalism requires a form of scientific realism. In general, there is indeed compatibility here, but there have also been some arguments given along similar lines to those in the previous paragraph. Sociologists of science have often viewed themselves as taking a properly naturalistic approach to science, in contrast with the philosophers' flights of fancy. So sociologists have sometimes argued that among the philosophical myths about science that we must abandon are myths about science's contact with reality.
So much for the possible tensions. The big one is between empiricism and realism. In the next section I will argue that the way to overcome this problem is via naturalistic ideas.
15.3 Empiricism Reformed
In this section I will describe a reformed version of empiricism. The argument will proceed in two steps. The first is a general philosophical discussion that has to do with epistemology in general. The second has to do with empiricism as a view about science.
As described above, empiricists (and many others) used to operate with a picture of the mind's access to the world that has been called the "veil of ideas" picture. The mind is seen as confined to its own sensations and thoughts, trying in vain to reach a hypothetical world beyond. Many philosophers now agree that this is a misleading picture. But it is easy to fall back into relatives of this view, both when thinking generally about the role of experience in guiding beliefs, and also in thinking about science. Philosophy of science has often hung onto enough of the old picture for trouble to arise. It is easy to fall back into a picture in which we distinguish two layers, or domains, in the world. One domain is accessible to us and familiar-the domain of experiences, or the domain of the observable. The other domain is inaccessible, mysterious, "theoretical," and problematic.
So how should we describe the role of experience? The right way to proceed is to cast empiricism within a naturalistic approach to philosophy. My version of this approach is influenced by the early-twentieth-century naturalism of John Dewey (19z9).
From the naturalistic point of view, humans are biological organisms embedded in a physical world that we evolved to deal with. All our livesincluding the most elaborate outgrowths of our social and intellectual lives-involve constant causal traffic and interaction with this world in which we are embedded. Our attempt to know about the world is just one aspect of our causal interaction with it; much of this interaction is more practical. Our perceptual mechanisms-eyes, ears, and so forth-are tools that we use to coordinate our dealings with the world. These mechanisms respond to physical stimuli caused by objects and events in our environments. From the inside, we can never establish with complete certainty what lies behind a particular sensory input. But looking at ourselves from "sideways-on," from the point of view taken by biology and psychology, we can establish regular principles concerning how our perceptual machinery responds to objects and events distant from us. We can work out how our perceptual machinery helps us to navigate the world.
So far this is not a point about science or even about human beings in particular. It is a claim about all the animals (and other organisms) that use perceptual mechanisms to adapt themselves to what is going on in their environments. But this point is enough to help us avoid some philosophical problems about our "access" to the world. We should not think in terms of two domains in reality, one accessible and one mysterious. We are biological systems embedded in a world containing objects of all sizes and at all different kinds of distance and remove from us. Our mechanisms of perception and action give us a variety of different kinds of contact with these objects. Our "access" to the world via thought and theory is really a complicated kind of causal interaction. This access to the world is constantly being expanded, as our technology improves. Parts of the world that must, at one time, be the subject of indirect and speculative inferences can later be much more directly observed, scanned, or assayed.
Think again about the problem of "underdetermination of theory by evidence" Empiricists have worried that there will always be rival alternative theories of the world that are equally compatible with our observations. Given this, how can we hope to have good knowledge of what the unobservable part of the world is really like? In thinking about this problem, return to the simpler case of perception itself. The same kind of issue arises here. Our perceptual mechanisms are used to form judgments about objects in the world around us, even though these mechanisms are only directly affected by stimuli like light and sound waves. In principle, there will always be alternative layouts of objects that could, in principle, give rise to the same stimuli affecting our senses. There is a kind of "underdetermination" here, as psychologists themselves often say. However, we can in fact make reliable judgments about what is around us, using perceptual mechanisms. And we can know that we are able to do this, by studying the operation of our perceptual mechanisms scientifically. In the case of perception, we can learn what kind of reliability we have in our attempts to know about the world.
The same sort of approach can be applied to inferences and modeling strategies in science itself. We can ask, What sort of reliability are we actually able to achieve, using different sorts of scientific reasoning and modelbuilding strategies? O
ver time, structures and objects in the world can move from being so inaccessible that only speculative modelbuilding can be applied to them, to being so accessible that their study is routine. Inferences about the genetic composition of an organism, for instance, have recently made this transition from being very indirect and cautious to being rather direct and routine, via DNA sequencing technologies. The constant shifting of these boundaries means that we can often go back and look at models developed when a structure was inaccessible. We can then ask, How well did we do? And more generally, Which approaches have tended to steer us toward good models, and which have tended to steer us toward bad ones? (Do scientific preferences for simple theories, for example, tend to lead to good choices?) To undertake this kind of investigation requires that we draw on work in the history of science. History tells us about actual cases where different approaches were tried and either succeeded or failed.
Let us now focus more closely on what makes science distinctive. Though humans all share their basic forms of contact with reality, as a consequence of their shared biological nature, there are huge differences in how different people and different intellectual cultures approach the problem of investigating and understanding the world. The fact that people all have their brains attached to sense organs at one end and behavioral mechanisms at the other end does not prevent them from disagreeing profoundly about the right way to learn about the world. One important point of disagreement is in how people handle the assessment of big ideas-big theories and explanatory hypotheses about the world-as opposed to how they handle everyday life. Maybe our shared biology is enough to make us all fairly empirical when we are trying to get food to eat and decide how to get home. But this does not apply to attempts to develop and justify explanatory theories about our overall place in the universe. Here we find sharp disagreements in approach, both within and across cultures.
Theory and Reality Page 31